Segmentation of Cerebrovascular Pathologies in Stroke Patients

with Spatial and Shape PriorsAdrian V. Dalca1, Ramesh Sridharan1, Lisa Cloonan2, Kaitlin M. Fitzpatrick2, Allison Kanakis2, Karen L. Furie3, Jonathan Rosand2, Ona Wu2, Mert Sabuncu4, Natalia S. Rost2, and Polina Golland1

Motivation

Spatial Model - Small Vessel Disease

Generative Model for T2-FLAIR

Segmentation Results

Cerebrovascular Pathologies • Small vessel disease predictor of stroke

outcome• Pathologies have similar intensities in

T2-FLAIR• Stroke has no consistent shape or

location pattern• Clinicians use spatial patterns in small

vessel disease to differentiate from stroke lesions

Current Methods• Segment hyperintensities via intensity

thresholds• Shape models used for other

segmentation tasks

Our Method• Model combines intensity and spatial

patterns• Inference to segment and differentiate

pathologies Inference & SegmentationSmall vessel disease prob.

Strokelesion prob.

Spatial model

Common Space• Register training subjects• Maps of small vessel disease

overlap(obtained manually)

Spatial Model • PCA on maps of small vessel disease• Components capture co-variation

e.g. bilateral periventricular symmetry

• Properties match those used by clinicians

• Can project estimated small vessel disease onto model

Segmentation• MAP estimate via Variational EM

inference

Initialize • Hyperintense voxels: small vessel

disease, stroke

E-Step Update• Estimate small vessel disease spatial

prior via regularized projection of small vessel disease

M-Step UpdatesSmall vessel disease and stroke probability maps• Estimate for class statistics

Include healthy tissue heterogeneity model

• Update tissue prior based on previous

MAP estimates and • Update posterior using new class

statistics, tissue prior, and neighbor agreement

Convergence leads to delineation of small vessel disease and stroke lesions simultaneously

1CSAIL, EECS, Massachusetts Institute of Technology 2Neurology, MGH, Harvard Medical School 3Neurology, RUH, Alpert Medical School. 4Martinos Center, Harvard Medical School

Manual vs Automatic segmentation• Good agreement of lesion outlines• Volume agreement (100 subj) • Volume agreement across manual

raters

Failure Cases• Bad registration• Large extent of small vessel

disease volumes

Small Vessel Disease Spatial Prior • Generate parameters

Tissue Priors for Class • Add MRF for spatial contiguity

Intensity Observations• Generated independently from Gaussian

Joint Probability

Goal: automatically segment and separate small vessel disease (leukoaraiosis) and stroke lesions in T2-FLAIR of stroke patients.

small vessel disease, stroke lesions

Stroke patients with pathologies

We model three tissue classes: small vessel disease, stroke, and healthy

small vessel disease

healthystroke

itera

tions

......

...

manual automatic failure cases